Oil & gas has always been a prediction industry
Long before artificial intelligence entered the conversation, oil & gas companies were building models of the future.
Engineers began developing mathematical approaches to estimating oil reserves more than a century ago. In 1945, J.J. Arps formalized the decline-curve equations that would remain a dominant approach to production forecasting for more than 70 years. Today, those methods sit alongside increasingly sophisticated statistical and machine-learning systems.
The reason is simple: few industries attach more value to understanding what is likely to happen next.
How much will a field produce?
When is critical equipment likely to fail?
Where will downtime occur?
Which interventions will have the greatest impact?
And where are the early signals that an operational or safety problem is developing?
These aren't peripheral questions. They influence production, reliability, safety, operating costs and the performance of assets worth billions of dollars.
The industry has spent decades building the data to answer them
Every well drilled, maintenance intervention completed and operating cycle recorded creates another piece of history.
Oil & gas operators have accumulated production histories, pressure and temperature readings, drilling parameters, workovers, equipment failures, inspection results, maintenance records, inventory, costs, weather conditions and commodity prices.
What makes that history particularly valuable isn't simply its volume. It's the repetition.
Thousands of wells experience variations of similar production challenges. Pumps and compressors move through recurring degradation and maintenance cycles. Facilities encounter operating conditions that have occurred elsewhere before. Every new well, failure, intervention and operating cycle creates another example from which models can learn.
That matters for performance. It also matters for safety.
Process safety and asset integrity are fundamentally concerned with preventing abnormal operating conditions or equipment failures from escalating into major incidents. The International Association of Oil & Gas Producers specifically emphasizes the importance of monitoring barriers and identifying the “weak signals” that can indicate future problems. IOGP
Those signals might appear in vibration, pressure, temperature, equipment condition, recurring alarms or other operating data. Identifying them earlier can give teams an opportunity to investigate, inspect or intervene before the situation becomes more serious. IOGP
Prediction in oil & gas, then, isn't simply about producing more efficiently. It can also help operators understand where risk is developing and make better-informed decisions about where attention is needed.
Better prediction is already creating measurable value
The industry doesn't need to be convinced that AI can create value.
ADNOC reported generating $500 million in additional value in 2023 through the deployment of more than 30 AI tools. Equinor reported $130 million in AI-driven value and savings in 2025 alone, taking its total since 2020 to $330 million.
Production forecasting provides one example. ADNOC has reported making production forecasts 90% more accurate, supporting more precise reservoir management and field development planning. Its AI program has also applied deep learning to pressure, temperature and vibration data to anticipate maintenance requirements before equipment failure.
The economics are particularly compelling in oil & gas because small improvements can have outsized consequences. A more accurate production forecast, additional uptime or earlier warning of equipment deterioration may each appear incremental in isolation. Across thousands of wells, extensive equipment fleets and assets worth billions of dollars, those improvements can compound into significant economic value.
The challenge is making prediction scale
Machine learning is already well established across the industry.
The challenge is often not whether another predictive model can be built. It is how much work is required to build it.
A conventional machine-learning project can require data scientists to prepare and transform data, engineer features, select an architecture, tune and validate the model, deploy it and then maintain it as conditions change.
For the industry's largest problems, the investment required to build a bespoke model can make sense. But operators make thousands of predictive decisions across wells, equipment, maintenance, planning and commercial operations. Individually, many may not justify a months-long machine-learning project. Collectively, they represent an enormous opportunity.
“Oil & gas companies don't have a shortage of problems worth predicting. The challenge is that historically you couldn't build and maintain a separate machine-learning system for every one of them. If you can change that equation, suddenly decades of operational data become useful across a much larger part of the operation.”
Oleg Zarakhani, Lead Data Scientist, Fundamental
That is perhaps the more interesting question for the industry's next phase of AI: not simply how to make a handful of existing models incrementally better, but how to make useful prediction practical across many more operational decisions.
What changes when the cost of prediction falls?
This is where a new generation of foundation models becomes interesting.
Large Tabular Models, or LTMs, are foundation models designed specifically to learn from structured data — the rows and columns found throughout operational databases and enterprise systems.
Instead of engineering a bespoke model from the ground up for every task, LTMs can learn relationships within new tabular datasets with substantially less manual feature engineering and tuning. Recent benchmark evaluations cited in our energy research show LTMs increasingly competing with and outperforming well-tuned traditional machine-learning approaches on tabular prediction tasks.
Early research is also beginning to show their relevance to energy and oil & gas.
A 2026 study applying the approach to natural-gas mixtures reported a mean error rate of 0.13% when predicting carbon-dioxide frosting temperature. Other research has applied tabular foundation models to solar forecasting and commodity markets, where their ability to capture nonlinear relationships and structural breaks could be particularly relevant in volatile conditions.
But the bigger opportunity isn't any single benchmark.
It's what happens when building another useful prediction becomes dramatically easier.
Lower that barrier and the possibilities expand. Production and reliability teams can test more hypotheses across more assets. Planning teams can evaluate more scenarios. Data science teams can support far more questions without building and maintaining a separate modeling project for every one.
That changes the threshold for where predictive AI becomes useful.
From a handful of AI projects to a predictive operation
This shift from a handful of high-value models to prediction across far more of the operation is the problem we're working on at Fundamental.
NEXUS is a Large Tabular Model built to make predictions directly from the structured data companies already have, automatically learning relationships and patterns without requiring the same level of manual feature engineering associated with traditional machine learning.
In one engagement with a Fortune 500 exploration and production company, NEXUS was evaluated against existing linear-regression models for energy demand and price forecasting across 13 regional markets. It delivered a 75% improvement in Mean Absolute Error and a 43% improvement in Root Mean Square Error, outperforming the existing approach in 70% of the regions evaluated and progressing into production.
But the larger opportunity isn't replacing the expertise that already exists inside oil & gas companies.
Reservoir engineers, production engineers, reliability teams, geoscientists and data scientists possess decades of domain knowledge. The opportunity is to give those teams the ability to put more of that knowledge - and more of the industry's extraordinary operational history to work.
For decades, operators have had to decide which problems are valuable enough to justify building a predictive model.
As that cost and complexity falls, the question changes:
What would you predict if you could make prediction practical across the operation?
Production. Reliability. Maintenance. Safety. Planning. Inventory. Commercial performance. Capital allocation.
For an industry that has spent more than a century trying to understand what happens next - and decades accumulating the data to do it - that could be one of the most consequential applications of AI yet.
The models will continue to improve. Compute will become cheaper. AI capabilities will become increasingly accessible.
But decades of production records, maintenance histories, drilling data, interventions and operating experience can't simply be downloaded.
Oil & gas already owns that advantage. The opportunity now is to put more of it to work.
